启动 HN：Relace（YC W23）– 快速可靠的代码生成模型
嘿，HN社区！我们是Preston和Eitan，我们正在构建Relace（[https://relace.ai](https://relace.ai)）。我们的目标是让构建代码代理变得简单且便宜。
以下是我们应用模型与整体文件编辑的一个示例：
[https://youtu.be/J0-oYyozUZw](https://youtu.be/J0-oYyozUZw)
构建可靠的代码代理是很困难的。在简单原型之外，任何在生产环境中进行代码生成的应用都会迅速遇到两个问题——如何可靠地应用差异，以及如何管理代码库的上下文？
我们专注于以数量级更低的价格和延迟解决这两个问题。
我们在二月份发布的第一个模型是快速应用模型——它以4300个标记的速度将代码片段与文件合并。在合并错误方面，它比Sonnet、Qwen、Llama或其他任何模型都更可靠。每个文件处理大约需要900毫秒，提供即时的用户体验，同时节省约40%的Claude 4输出标记。
我们的第二个模型专注于检索。对于基于情感编码和企业代码库，仅检索与用户请求相关的文件可以节省最新技术的输入标记成本，并减少代码代理查看文件的次数。我们的重排序器（评估如下）可以在约1-2秒内扫描百万行代码库，而我们的嵌入模型在Typescript/React代码库的检索评估中优于任何其他嵌入模型。
构建编码代理有很多不同的方法，但能够可靠地编辑代码并检索代码库中最相关的部分将是一个基础性问题。我们很高兴能够构建出更易于数百万用户使用的方式，而这些用户不想在Claude上花费大量资金。
这些模型在生产中每周使用数百万次。如果你使用过Lovable、Create.xyz、Magic Patterns、Codebuff或Tempo Labs，那么你就已经在使用我们的技术了！
以下是试用的链接：[https://app.relace.ai](https://app.relace.ai)，这是我们的文档：[https://docs.relace.ai](https://docs.relace.ai)。
我们已经为每个人开放了原型设计的免费访问，限制应该足够用于个人编码和构建小项目（如果不够请纠正我们）。我们直接与像Continue.dev这样的开源IDE集成。请试用我们的服务，我们期待听到你的反馈！
查看原文
Hey HN community! We're Preston and Eitan, and we're building Relace (<a href="https://relace.ai">https://relace.ai</a>). We're trying to make building code agents easy and cheap.<p>Here’s an example of our apply model vs. whole file edits:
<a href="https://youtu.be/J0-oYyozUZw" rel="nofollow">https://youtu.be/J0-oYyozUZw</a><p>Building reliable code agents is hard. Beyond simple prototypes, any app with code generation in production quickly runs into two problems -- how do you reliably apply diffs, and how do you manage codebase context?<p>We're focused on solving these two problems at order-of-magnitude lower price and latency.<p>Our first model that we released, in February, is the Fast Apply model -- it merges code snippets with files at 4300 tok/s. It is more reliable (in terms of merge errors) than Sonnet, Qwen, Llama, or any other model at this task. Each file takes ~900ms and gives an instantaneous user experience, as well as saving ~40% on Claude 4 output tokens.<p>Our second model focuses on retrieval. For both vibe-coded and enterprise codebases, retrieving only the files relevant to a user request saves both on SoTA input token cost and reduces the number of times code agents need to view files. Our reranker (evals below) can scan a million-line codebase in ~1-2s, and our embedding model outperforms any other embedding model for retrieval as evaluated on a corpus of Typescript/React repositories.<p>There are many different ways to build coding agents, but being able to edit code reliably and retrieve the most relevant parts of the codebase is going to be a foundational issue. We're excited to be building ways to make it more accessible to millions of users who don't want to spend $$$ on Claude.<p>These models are used in production, millions of times per week. If you've used Lovable, Create.xyz, Magic Patterns, Codebuff, Tempo Labs then you've used us!<p>Here's a link to try it out: <a href="https://app.relace.ai">https://app.relace.ai</a>, and here are our docs: <a href="https://docs.relace.ai">https://docs.relace.ai</a>.<p>We've opened up free access for prototyping on our website to everyone, and the limits should be enough for personal coding use and building small projects (correct us if it’s not). We integrate directly with Open-Source IDE's like Continue.dev. Please try us out, we'd love to hear your feedback!